A Fuzzy Ontology and Its Application to News Summarization

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 35, NO. 5, OCTOBER 2005 859

A Fuzzy Ontology and Its Applicationto News SummarizationChang-Shing Lee, Zhi-Wei Jian, and Lin-Kai Huang

Abstract—In this paper, a fuzzy ontology and its application tonews summarization are presented. The fuzzy ontology with fuzzyconcepts is an extension of the domain ontology with crisp concepts.It is more suitable to describe the domain knowledge than domainontology for solving the uncertainty reasoning problems. First,the domain ontology with various events of news is predefinedby domain experts. The document preprocessing mechanism willgenerate the meaningful terms based on the news corpus andthe Chinese news dictionary defined by the domain expert. Then,the meaningful terms will be classified according to the eventsof the news by the term classifier. The fuzzy inference mechanismwill generate the membership degrees for each fuzzy concept ofthe fuzzy ontology. Every fuzzy concept has a set of membershipdegrees associated with various events of the domain ontology.In addition, a news agent based on the fuzzy ontology is alsodeveloped for news summarization. The news agent contains fivemodules, including a retrieval agent, a document preprocessingmechanism, a sentence path extractor, a sentence generator, and asentence filter to perform news summarization. Furthermore, weconstruct an experimental website to test the proposed approach.The experimental results show that the news agent based on thefuzzy ontology can effectively operate for news summarization.

Index Terms—Chinese natural language processing, fuzzy infer-ence, news agent, news summarization, ontology.

I. INTRODUCTION

AN ontology is a formal conceptualization of a real world,and it can share a common understanding of this real world

[11]. With the support of the ontology, both user and systemcan communicate with each other by the shared and commonunderstanding of a domain [16]. There are many ontologicalapplications that have been presented in various domains. Forexample, Embley et al. [3] present a method of extracting infor-mation from unstructured documents based on an applicationontology. Alani et al. [1] propose the Artequakt that automat-ically extracts knowledge about artists from the web based onan ontology. It can generate biographies that tailor to a user’sinterests and requirements. Navigli et al. [14] propose the On-toLearn with ontology learning capability to extract relevantdomain terms from a corpus of text. OntoSeek [4] is a systemdesigned for content-based information retrieval. It combines

Manuscript received June 11, 2004; revised October 10, 2004 and January9, 2005. This work was supported in part by the National Science Council ofTaiwan (R.O.C.) under Grant NSC-93-2213-E-309-003 and by the Ministry ofEconomic Affairs in Taiwan under Grant 93-EC-17-A-02-S1-029. This paperwas recommended by Associate Editor Hisao Ishibuchi.

C.-S. Lee and L.-K. Huang are with the Department of Information Man-agement, Chang Jung Christian University, Tainan, Taiwan, R.O.C. (e-mail:[email protected]; [email protected]).

Z.-W. Jian is with the Department of Computer Science and Information En-gineering, National Cheng Kung University, Tainan, Taiwan, R.O.C.

Digital Object Identifier 10.1109/TSMCB.2005.845032

an ontology-driven content-matching mechanism with moder-ately expressive representation formalism. Handschuh et al.[6] provide a framework, known as S-CREAM, that allowsfor creation of metadata and is trainable for a specific domain.Ont-O-Mat is the reference implementation of the S-CREAMframework. It provides a plugin interface for extensions forfurther advancements, e.g., collaborative metadata creation orintegrated ontology editing and evolution. Vargas-Vera et al.[17] present an annotation tool, called MnM, which providesboth automated and semi-automated support for annotatingweb pages with semantic contents. MnM integrates a webbrowser with an ontology editor and provides open applica-tion programming interfaces (APIs) to link to ontology serversand for integrating information extraction tools.

The goal of text summarization is to take the abstract fromthe extracted content and present the most important messagefor the user in a condensed form. McKeown et al. [13] de-velop a composite summarization system that uses differentsummarization strategies dependent on the type of documentsin the input set. Mani [12] introduces two text summary ap-proaches, including topic-focused and generic summary. Lamet al. [9] propose the Financial Information Digest System(FIDS) to summarize online financial news articles automati-cally. The FIDS can integrate the information from differentarticles by conducting automatic content-based classificationand information item extraction. Halteren [5] develops a newtechnique to find out several feature sets through the set ofstyle makers and extract summarization sentences from thedocument, according to the found feature sets for users. Leeet al. [8] propose an ontology-based fuzzy event extractionagent for Chinese news summarization. The summarizationagent can generate a sentence set for each Chinese news, butthe results need to be improved to suit for various Chinesenews websites. Riloff [15] proposes the AutoSlog system thatcan automatically create domain-specific dictionaries for in-formation extraction by an appropriate training corpus.

In this paper, we present a fuzzy ontology and apply it tonews summarization. The fuzzy ontology is an extension of thedomain ontology that is more suitable to describe the domainknowledge for solving the uncertainty reasoning problems. Inaddition, a news agent based on the fuzzy ontology is also de-veloped for news summarization. The remainder of this paper isstructured as follows: Section II describes the definition of thefuzzy ontology. Section III introduces a fuzzy inference mech-anism for constructing the fuzzy ontology. Section IV proposesa news agent based on the fuzzy ontology for news summariza-tion. Section V shows some experimental results for Chinesenews summarization. Section VI presents our conclusions andthe future work.

1083-4419/$20.00 © 2005 IEEE

860 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 35, NO. 5, OCTOBER 2005

(a)

(b)

Fig. 1. (a). Structure of the domain ontology. (b) Example of the domain ontology for weather news domain.

LEE et al.: FUZZY ONTOLOGY AND ITS APPLICATION TO NEWS SUMMARIZATION 861

(c)

Fig. 1. (Continued) (c) Corresponding Chinese version of Fig. 1(b).

II. DEFINITION OF THE FUZZY ONTOLOGY

In our previous work [8], we have extended the three-lay-ered object-oriented ontology architecture to the four-layeredmodel for domain ontology representation. The performancesof the experimental results [8] are effective, so we still adoptthe four-layered model in this paper. Fig. 1(a) and (b) show thefour-layered structure of the domain ontology and an exampleof weather news domain ontology, respectively. In Fig. 1(b),the domain name is “Weather news,” and it consists of severalcategories such as “Weather report,” “Weather encyclopedia,”and “Astronomy.” In addition, there are several events definedin this domain ontology. For example, the weather events “ :Typhoon,” “ : Cold current,” and “ : Rain” are related tothe categories “Weather report,” “Weather encyclopedia,” and“Astronomy.” Each event comprises several concepts, such as“Central Weather Bureau,” “Present Time,” “Temperature,” and“Weather.” Fig. 1(c) shows the corresponding Chinese versionof Fig. 1(b).

In this section, we further present a fuzzy ontology for newssummarization. Because the fuzzy ontology is an extension ofthe domain ontology, we briefly introduce the domain ontologyas follows. There are many different descriptions of the domainontology for various applications. Now, we give a formal defi-nition of the domain ontology for news summarization.

[Definition 1] Domain Ontology: A domain ontology de-fines a set of representational terms that we call concepts.Inter-relationships among these concepts describe a targetworld [11]. There are four layers, including domain layer,category layer, event layer, and class layer, defined in thedomain ontology [8]. The domain layer represents the domainname of an ontology and consists of various categories definedby domain experts. Each category is composed of the eventset , which is derived from the news corpusby domain experts. Every event comprises several conceptsof class layer. In class layer, each concept contains a conceptname , an attribute set , and an operationset for an application domain. There are


(a)

(b)

Fig. 2. (a) Structure of the fuzzy ontology. (b) Example of the fuzzy ontology for weather news domain.


(c)

Fig. 2. (Continued). (c) Corresponding Chinese version of Fig. 2(b).

three kinds of relationship, including generalization, aggrega-tion, and association, in the domain ontology. The relationshipbetween a domain and its corresponding category is generaliza-tion that represents “is-kind-of” relationship. The relationshipbetween each category and its corresponding events is aggre-gation. The aggregation denotes “is-part-of” relationship. Theassociation represents a semantic relationship between conceptsin class layer.

In this paper, we extend the domain ontology to be the fuzzyontology by embedding a set of membership degrees in eachconcept of the domain ontology and adding two fuzzy relation-ships among the fuzzy concepts. The concept with the member-ship degrees is called fuzzy concept. Now, we give the defini-tions of fuzzy concept, fuzzy relationship, and fuzzy ontology asfollows.

[Definition 2] Fuzzy Concept: A fuzzy concept is a refinedconcept derived from a domain ontology. It is a refinement by

embedding a set of membership degrees associated with a setof the news events in the concept of the domain ontology. If adomain ontology has an event set and a con-cept , then we can refine the into the fuzzy concept and de-note the fuzzy concept as withan attribute set and an operation set

, where represents the member-ship degree of for event . Besides, the attribute andthe operation denote the th attribute and th operationof , respectively.

[Definition 3] Fuzzy Relationship: There are two kinds offuzzy relationships, including Location_Narrower_Relation-ship (LNR) andLocation_Broader_Relationship (LBR), definedfor news summarization. If there exist two fuzzy concepts (loca-tion and location and ), then we denote the fuzzyrelationships LNR andLBR as and ,respectively.


Fig. 3. Process of the fuzzy ontology construction.

[Definition 4] Fuzzy Ontology: A fuzzy ontology is anextended domain ontology with fuzzy concepts and fuzzyrelationships.

Fig. 2(a) and (b) show the structure of the fuzzy ontology andan example of weather news fuzzy ontology, respectively.

In Fig. 2(b), the domain name is “Weather news,” and it con-sists of several categories such as “Weather report,” “Weatherencyclopedia,” and “Astronomy.” In addition, there are sev-eral events defined in this domain ontology. For example, theweather events “ : Typhoon,” “ : Cold current,” and “ :Rain” are related to the categories “Weather report,” “Weatherencyclopedia,” and “Astronomy,” respectively. Each eventcomprises several fuzzy concepts, such as “{Central WeatherBureau; 0.99, 0.85, 1.0},” “{Present Time; 0.76, 0.87, 0.94},”“{Temperature; 0.04, 1.0, 0.64},” and “{Weather; 0.5, 0.94,0.84}.” For example, the membership degrees of the fuzzyconcept “Temperature” for weather events “ : Typhoon,”“ : Cold current,” and “ : Rain” are 0.04, 1.0, and 0.64,respectively. Fig. 2(c) shows the corresponding Chinese versionof Fig. 2(b).

III. FUZZY INFERENCE MECHANISM FOR FUZZY

ONTOLOGY CONSTRUCTION

In this section, we apply a fuzzy inference mechanism to con-struct the fuzzy ontology. Fig. 3 shows the process of the fuzzyontology construction.

In Fig. 3, the news domain ontology with various events ispredefined by domain experts. The document preprocessingmechanism will generate the meaningful terms based on thenews corpus produced by the retrieval agent and the Chinesenews dictionary defined by domain experts. Then, the termclassifier will classify the meaningful terms according to theevents of the news. The fuzzy inference mechanism will generatethe membership degrees for each fuzzy concept of the fuzzyontology. Every fuzzy concept has a set of membership degreesassociated with various events of the domain ontology. Now,we briefly describe the retrieval agent, document preprocessingmechanism, and term classifier as follows.

A. Retrieval Agent, Document Preprocessing Mechanism,and Term Classifier

First, the retrieval agent periodically retrieves the news fromone of the largest news Websites in Taiwan (http://www.china-times.com). The retrieved news will be stored into a news corpusfor further processing. The document preprocessing mechanismconsists of a part-of-speech (POS) tagger [19] provided by Chi-

nese Knowledge Information Processing (CKIP) [19] Group inTaiwan and a term filter to get the meaningful term set includingnouns and verbs. Table I shows a part of POS tags.

The term filter will preserve the meaningful terms accordingto the POS tags [8] selected by the domain experts based on theChinese news dictionary [20] and the term frequencies of thenews corpus. In addition, the meaningful terms will be classifiedaccording to the events of the news by the term classifier. Now,we describe the algorithm of term classifier as follows.

The algorithm of Term ClassifierInput:A domain ontology and all meaningfulterms of news generated by document pre-processing mechanism.Output:Classified meaningful term sets.Method:1) Parse and store all terms of var-ious events of the domain ontology into

denotes theterm set of event in the domain ontology.

2) . denotesthe term set of the output event .3) For to denotesthe number of news .3.1) For to denotes thenumber of all terms in the th news .3.1.1) For to de-

notes the number of events in the domainontology .

3.1.1.1) If then de-notes the th term of input terms in theth news .

denotes the numberof terms in the event .3.2) For to3.2.1) If is the largest, then3.2.1.1) For to

.4) End.

B. Fuzzy Inference Mechanism

In our previous work [8], we have presented a five-layerFuzzy Inference Agent (FIA) [7], [10] for news event ontology


TABLE IPART-OF-SPEECH TAGS OF CHINESE TERMS DERIVED FROM CKIP

extraction. In this paper, we modify the FIA and extend it to aseven-layer Fuzzy Inference Mechanism (FIM) for creating thefuzzy ontology. Layers 6 and 7 of FIM are utilized to computeand integrate the fuzzy membership degrees and then generatethe fuzzy concepts for each event of the news. Fig. 4 showsthe proposed FIM. There are seven layers, including inputlinguistic layer, input term layer, rule layer, output term layer,output linguistic layer, summation layer, and integration layerin this architecture.

Now, we describe each layer of FIM for event in detail.1) Layer 1 (Input Linguistic Layer): The input linguistic

layer contains two kinds of sets, including term set ofevents provided by term classifier and concept set of domainontology. The nodes in first layer just directly transmit inputvalues to next layer. The input vectors are the term set andthe concept set retrieved from term classifier and the domainontology, respectively. The term set derived from term classifieris , and the concept set of the domainontology is .

2) Layer 2 (Input Term Layer): The input term layer per-forms the membership functions to compute the membership

degrees for all terms derived from the retrieved news and thedomain ontology. In this paper, the fuzzy number of L-R type[18] is adopted for the fuzzy variables and formulated by thefollowing equation:

for

for(1)

where , and can be repre-sented as a triplet .

There are three input fuzzy variables, including Term Part-of-Speech (POS), Term Word (TW) similarity, and Semantic Dis-tance (SD) similarity, considered in input term layer for eachterm’s property. The structure of input term node contains twoparts, including Fuzzy Variable Part and Linguistic Term Part, toperform the membership degree computing. The Fuzzy VariablePart will extract the input values fromand for the fuzzy variable and send the ex-tracted values to the Linguistic Term Part. In addition, the Lin-guistic Term Part operates to calculate the membership degreewith respect to the input values.


Fig. 4. Architecture of fuzzy inference mechanism.

(a)

(b)

Fig. 5. (a). Structure of input term node. (b) Example of input term node forTerm POS fuzzy variable.

Fig. 5(a) and (b) show the structure of the input term node andan example of the input term node for Term POS fuzzy variable

with three linguistic terms, including High, Median, and Low,respectively.

In Fig. 6, each node of the tagging tree represents apart-of-speech of a term enumerated in the Table I. We willutilize the length of the path between every two nodes tocompute the POS similarity for each term pair. The path lengthof every two terms defined by the tagging tree is bounded in theinterval [0, 7]. For example, the two terms “ (Typhoon, Na)”and “ (Taiwan, Nc)” with their POS are “Na” and “Nc,”respectively; hence, the path length of the two terms is 2Na Nc . There are three linguistic terms, in-

cluding POS_High, POS_Median, and POS_Low defined forPOS similarity. Fig. 7 shows the membership functions ofthe fuzzy sets for POS similarity. The membership degreeis usually a value in the range [0, 1], where “1” denotes afull membership, and “0” denotes no membership. Noticethat domain experts predefine the membership functions ofPOS_High, POS_Median, and POS_Low. In this paper, thefuzzy numbers for POS_High, POS_Median, and POS_Loware ,

, and, respectively. Table II shows the parameter values

of the fuzzy numbers for POS similarity defined by domainexperts.

The second fuzzy variable proposed for computing therelation strength of any term pair is Term Word (TW) similarity.In this section, we will compute the number of the identicalwords that each term pair has to get TW similarity. Inaddition, because CKIP defines that the maximum numberof words for any terms is 6, the bound of the number ofthe same word for any term pair is [0, 6]. For example,


Fig. 6. Tagging tree derived from CKIP.

Fig. 7. Fuzzy sets of POS similarity for each term pair.

the term “ (Government Information Office)”with six Chinese words is aterm with the maximum number of words. Next,the two terms “ (A medium typhoon)”and “ (A light typhoon)” have three of thesame Chinese words and hence, the numberof the same words is 3. Table II shows the parameter values ofthe fuzzy numbers TW_Low, TW_Median, and TW_High forTW similarity fuzzy variable defined by domain experts.

The last fuzzy variable considered for input term pair isSemantic Distance (SD) similarity. In the SD similarity com-putation, the domain ontology will be defined as five layers,including the who layer, the when layer, the what layer, the

where layer, and the how layer, by domain experts. For example,the two terms “ (Airflow)” and “ (Temperature)” inFig. 1(b) are both located in the same what layer; hence, thesemantic distance is 2, shown at the bottom of the page. Anexample of the typhoon event ontology with the five layers isshown in Fig. 1(b). In this paper, we compute the semanticdistance in the same layer of domain ontology for any termpair. Table II shows the parameter values of the fuzzy numbersfor SD similarity fuzzy variable defined by domain experts.

Each fuzzy variable of the input term layer appearing in thepremise part is represented with a condition node. Each of theoutputs of the condition node is connected to rule nodes in thethird layer to constitute a condition specified in some rules. This


TABLE IIPARAMETER VALUES OF THE FUZZY NUMBERS

layer performs the first inference step to compute matching de-grees. If the input vector of this layer is

then it will be transferred as follows:

(2)

where represents the term pair , is the thterm of the news for event 1, is the th concept of the domainontology, is the part-of-speech path length of eachterm pair, represents the same word of each term pair,and is the semantic distance of each term pair. Then,the output vector can be written as follows:

(3)

TABLE IIIFUZZY INFERENCE RULES FOR FIM

where is the membership degree of the th linguisticterm in the fuzzy variable of POS, is the membershipdegree of the th linguistic term in the fuzzy variable of TW, and

is the membership degree of the th linguistic term inthe fuzzy variable of SD.

3) Layer 3 (Rule Layer): The third layer is called the rulelayer, where each node is a rule node to represent a fuzzyrule. The links in this layer are used to perform preconditionmatching of fuzzy logic rules. Hence, the rule nodes shouldperform the fuzzy AND operation [10], and the outputs will belinked with associated linguistic node in the fourth layer. In thispaper, we use the algebraic product operation to compute thematching degree. In our model, the rules are defined by domainexpert’s knowledge previously, and we show them in Table III.Equation (4) shows the precondition matching degree of rulenode 1 as follows:

(4)

4) Layer 4 (Output Term Layer): The output term layerperforms the fuzzy OR operation to integrate the fired rulesthat have the same consequences. The fuzzy variable definedin output layer is Terms Relation Strength (TRS). There arefive linguistic terms, including TRS_Very_Low, TRS_Low,TRS_Median, TRS_High, and TRS_Very_High in TRS. Table IIshows the parameter values of the fuzzy number for TRS fuzzyvariable defined by domain experts.


For example, if there are rules inferring the same conse-quences with the th output linguistic term Low, and the corre-sponding gravity of linguistic term Low is , then the outputof this layer for term pair is as follows:

(5)

5) Layer 5 (Output Linguistic Layer): The output linguisticlayer performs the defuzzification process to get the TRS of theterm pair. In this paper, the center of area (COA) method [10]is adopted to carry out the defuzzified process. Finally, the TRSvalue for term pair of event is as follows:

(6)

where is the number of rule nodes, is the number of linguisticterms of output fuzzy variable, and is the gravity of thoutput linguistic term associated with the th rule node. There-fore, in our case, the values of and are 27 and 5, respectively.

6) Layer 6 (Summation Layer): The summation layer per-forms the summation process for computing the TRS values ofthe specific concept of event 1. The input vector of this layeris as follows:

(7)

where is the TRS value of th term and th conceptfor event 1. The summation process is as follows:

for all (8)

7) Layer 7 (Integration Layer): Finally, the integration layerwill integrate the membership degrees into the fuzzy ontology.The integration node of this layer will integrate the membershipdegrees of the concept that belongs to all events of the do-main ontology, that is, . Finally,the fuzzy ontology will be represented as follows:

...

(9)

where the output vector de-notes the membership degree set of th fuzzy concept.

Fig. 8. Process of news agent for news summarization.

IV. APPLICATION OF FUZZY ONTOLOGY TO CHINESE

NEWS SUMMARIZATION

In this section, we propose a news agent based on the fuzzyontology for news summarization. In particular, we focus onChinese news processing in this paper.

Fig. 8 shows the process of news agent for news summariza-tion. The retrieval agent and document preprocessing mecha-nism perform the same tasks as we described in Section III.Moreover, the Sentence Path Extractor (SPE) will extract a setof possible sentence paths from the fuzzy ontology and send thetemplate set to Sentence Generator (SG). The SG will producea set of sentences from the class layer of the fuzzy ontology.Finally, the Sentence Filter (SF) will filter noisy sentences andcreate a set of summarized sentences.

A. Sentence Path Extractor

The Sentence Path Extractor uses the Depth-First-Search al-gorithm [2] to look for possible sentence paths from the fuzzyontology. Now, we describe the algorithm as follows.

The algorithm of Sentence Path ExtractorInput:All classified meaningful terms generatedby document preprocessing mechanism termsof the news and all fuzzy concepts of thefuzzy ontology.Output:A set of sentence paths.Method:1) . repre-sents a temporary fuzzy ontology .2) For all terms generated by documentpreprocessing mechanism, and or

is a re-lation between and2.1) If then2.1.1) If thenJoin to .2.1.2) Join to .2.2) If thenJoin to .

3) For all concepts of3.1) For toThe denotes the maximum indegree of thefuzzy concept in


3.1.1) For to Thedenotes the number of the fuzzy concept in

3.1.1.1) If (indegree of ) and(outdegree of ) then

Join into de-notes a set of the fuzzy concept with in-degree .4) . de-notes a set of the sentence path .5) For to5.1) For to denotesthe element number of the concept set

5.1.1) .5.1.2) While isn’t the leaf node of

5.1.2.1) .5.1.2.2) .5.1.3) For to de-

notes the number of the child nodes of5.1.3.1) .5.1.3.2) Go to While statement.

6) End.

The result of Sentence Path Extractor is denoted as follows:

Sentence Path P: [Concept Name] [Concept Name][Concept Name]

Fig. 9 shows an example of a temporary fuzzy ontology TFOextracted from Fig. 2(b). There are two possible sentence pathsextracted by SPE shown as follows:

Sentence Path 1:Corresponding English Version:

“[Central weather bureau] [Future time] [Ty-phoon] [Taiwan] ”

and

Sentence Path 2:Corresponding English Version:

“[Central weather bureau] [Future time][Rain]”.

B. Sentence Generator

The Sentence Generator generates a set of sentences basedon the temporary fuzzy ontology. The algorithm of SentenceGenerator is as follows:

The algorithm of Sentence GeneratorInput:A temporary fuzzy ontology TFO and a setof sentence paths.Output:A set of sentences.Method:1) For to de-notes the number of sentence paths in thesentence path set .

(a)

(b)

Fig. 9. (a) Temporary fuzzy ontology TFO. (b) Corresponding English versionof Fig. 9(a).

1.1) Compute the number of con-cepts in the sentence path .

denotes the th sentence path of .1.2) For to1.2.1) . denotes the pos-

sibility of the sentence .1.2.2) If then denotes

the th sentence1.2.2.1) Get all information of conceptfrom the temporary fuzzy ontology.


(a)

(b)

(c)

Fig. 10. Average precision of typhoon event evaluated by three domainexperts. (a) Results of 2002 typhoon news. (b) Results of 2003 typhoon news.(c) Results of 2004 typhoon news.

1.2.2.2) . de-notes the membership degree of th conceptof in th event1.2.3) Else1.2.3.1) Get all information of conceptfrom the temporary fuzzy ontology.1.2.3.2) Get a relation between

and from the temporary fuzzy on-tology.

1.2.3.3) Use the relation toconnect with and store them into.1.2.3.4) .

1.3) If of the sentence .

(a)

(b)

(c)

Fig. 11. Average precision of cold current event evaluated by three domainexperts. (a) Results of 2002 cold current news. (b) Results of 2003 cold currentnews. (c) Results of 2004 cold current news.

Combine the sentence with andstore them into the sentence set.2) End.

The result of Sentence Generator is denoted as follows:

Sentence : [Concept Name, Attribute_value, Oper-ation] Relation [Concept Name, Attribute_value,Operation] Relation [Concept Name, At-tribute_value, Operation] (Possibility )

By the algorithm of Sentence Generator, the news agent willgenerate the following sentences:

Sentence 1: This is the sentence shown at the bottomof the page.


(a)

(b)

(c)

Fig. 12. Average precision of rain event evaluated by three domain experts.(a) Results of 2002 rain news. (b) Results of 2003 rain news. (c) Results of2004 rain news.

Corresponding English Version:“[Central weather bureau, Null, Null] Expect

[Future time, Tomorrow, Null] Has [Typhoon, Amedium typhoon, Null] Attack [Taiwan, Taiwan,Null] (Possibility 0.78)”English Meaning:

The central weather bureau expects that the typhoonwill attack Taiwan tomorrow. (Possibility 0.78).

Sentence 2: This is the sentence shown at the bottomof the page.

(a)

(b)

(c)

Fig. 13. Average recall of typhoon event evaluated by three domain experts.(a) Results of 2002 typhoon news. (b) Results of 2003 typhoon news. (c) Resultsof 2004 typhoon news.

Corresponding English Version:“[Central weather bureau, Null, Null] Expect

[Future time, Tomorrow, Null] Will [Rain, Null,Occur] (Possibility 0.84)”English Meaning:

The central weather bureau expects that it will rainin Taiwan tomorrow. (Possibility 0.84).

C. Sentence Filter

The Sentence Filter is used to filter the redundant sentences.It will generate a brief sentence set. The Sentence Filter will


(a)

(b)

(c)

Fig. 14. Average recall of cold current event evaluated by three domainexperts. (a) Results of 2002 cold current news. (b) Results of 2003 cold currentnews. (c) Results of 2004 cold current news.

combine the common sentences and transfer them into semanticsentences. The algorithm of Sentence Filter is as follows:

The algorithm of Sentence FilterInput:A sentence set generated by Sentence Gen-erator and the news retrieved by the re-trieval agent.Output:The summary results of the news.Method:1) Sort the length of each sentence of thesentence set descending in order.2) For to denotes the numberof the sentences in the sentence set .

(a)

(b)

(c)

Fig. 15. Average recall of rain event evaluated by three domain experts.(a) Results of 2002 rain news. (b) Results of 2003 rain news. (c) Results of2004 rain news.

2.1) For to2.1.1) .2.1.2) .2.1.3) If or .2.1.3.1) Divide into . de-

notes a set with the concepts of the sen-tence .

2.1.3.2) Divide into . de-notes a set with the concepts of the sen-tence .

2.1.3.3) If thenDelete from .

3) Add the date of the retrieved news to. denotes the summary results

of the news .


(a)

(b)

(c)

Fig. 16. Average precision curves of fuzzy ontology-based agent and OFEEagent for typhoon event. (a) Results of 2002 typhoon news. (b) Results of 2003typhoon news. (c) Results of 2004 typhoon news.

4) Add the event title of the retrievednews to .5) For to5.1) .

denotes the semantic sentence .5.2) Divide into .

denotes a set with the result of thesentence in the sentence set .5.3) For to

denotes thenumber of .

. denotes the thconcepts or relations of5.4) Add to .

(a)

(b)

(c)

Fig. 17. Average precision curves of fuzzy ontology-based agent and OFEEagent for cold current event. (a) Results of 2002 cold current news. (b) Resultsof 2003 cold current news. (c) Results of 2004 cold current news.

6) Compute the compression rate based onand the news, and add it to .

7) Store into the summary repository.8) End.

For example, the following sentences and will becombined and transferred to Semantic Sentence SS by SentenceFilter as follows:

Sentence: This sentence is shown at the bottom of thepage.


(a)

(b)

(c)

Fig. 18. Average precision curves of fuzzy ontology-based agent and OFEEagent for rain event. (a) Results of 2002 rain news. (b) Results of 2003 rain news.(c) Results of 2004 rain news.

Corresponding English Version:: “[Central weather bureau, Null, Null] Expect

[Future time, Tomorrow, Null] Has [Typhoon,Null, Null] Attack [Taiwan, Null, Null] (Possi-bility 0.78)”

: “[Typhoon, Null, Null] Attack [Taiwan,Null, Null] (Possibility 1.0)”English Meaning:

: The central weather bureau expects that the ty-phoon will attack Taiwan tomorrow (Possibility 0.78).

: The typhoon will attack Taiwan (Possibility 1.0).Result of the semantic sentence:

Corresponding English Version:SS: The central weather bureau expects that the ty-

phoon will attack Taiwan tomorrow (Possibility 0.78).

(a)

(b)

(c)

Fig. 19. Average recall curves of fuzzy ontology-based agent and OFEEagent for typhoon event. (a) Results of 2002 typhoon news. (b) Results of 2003typhoon news. (c) Results of 2004 typhoon news.

V. EXPERIMENTAL RESULTS

We have constructed an experimental website (http://210.70.151.160/Experimental_Results/) to test the perfor-mance of the proposed approach. The website is also locatedat “Decision Support and Artificial Intelligent Lab.” of ChangJung University in Taiwan, and implemented with JAVA onLinux operating system. The experimental Chinese news isretrieved from one of the largest news websites in Taiwan(http://www.chinatimes.com). There are some evaluationmodels for document summarization. For example, the DUCmultidocument summarization evaluation involved 30 docu-ment sets [13]. For each test data set and each target summarysize, one automatically generated summary was submitted fromeach participating site, and one gold-standard summary wascreated by humans. Comparisons for each data set and targetsummary size involved the human-created summary versus thesummaries automatically produced by competing systems [13].


(a)

(b)

(c)

Fig. 20. Average recall curves of fuzzy ontology-based agent and OFEE agentfor cold current event. (a) Results of 2002 cold current news. (b) Results of 2003cold current news. (c) Results of 2004 cold current news.

In this paper, we collected the Chinese news of three weatherevents, including “Typhoon event,” “Cold current event,” and“Rain event,” from Chinatimes website in 2002, 2003, and 2004.Twenty pieces of news are adopted for each event to test theperformance of our approach. Hence, the total news numberof three events is 180. There are three domain experts partic-ipating in the experiment for evaluating the performance of our

(a)

(b)

(c)

Fig. 21. Average recall curves of fuzzy ontology-based agent and OFEE agentfor rain event. (a) Results of 2002 rain news. (b) Results of 2003 rain news.(c) Results of 2004 rain news.

approach. We adopt the performance measures Precision andRecall in this experiment. The formulas of Precision and Recallmeasures utilized in this paper are as in (10) and (11), shown atthe bottom of the page.

Figs. 10–12 show the average precision of “Typhoon event,”“Cold current event,” and “Rain event” evaluated by the threedomain experts, respectively. Figs. 13–15 show the averagerecall of “Typhoon event,” “Cold current event,” and “Rain

Precision

The number of relevant sentences between gold standard summary andautomatically generated summary

The number of sentences in gold standard summary(10)

Recall

The number of relevant sentences between gold standard summary andautomatically generated summary

The number of sentences in automatically generated summary(11)


(a)

(b)

Fig. 22. (a). News of typhoon event retrieved from ChinaTimes website. (b) The corresponding English version of Fig. 22(a).

event” evaluated by the three domain experts, respectively.These results show that the average precision and the averagerecall are stable when increasing the number of documents.

Next, we test the performance of the news agent based onfuzzy ontology by comparing with our previous work on theOntology-based Fuzzy Event Extraction (OFEE) agent [8].Figs. 16–19 show the compared curves for the average pre-cision measure, respectively. Figs. 20–22 show the comparedresults for the average recall measure, respectively. By theexperimental results, we observe that the news agent with fuzzyontology can achieve better summarized results than the OFEEagent.

Finally, we illustrate three examples for Chinese weathernews summarization. Figs. 22(a) and 23(a) show the news oftyphoon event that retrieved from the ChinaTimes website andthe summarization result, respectively. Figs. 22(b) and 23(b)shows the corresponding English version of Figs. 22(a) and23(a), respectively.

Notice that the summarization sentences for typhoon eventincluding “S1 (Typhoon has moved away

from Taiwan),” “S2 (Taiwan has showers),” and“S3 (Central Weather Bureau forecaststhat Taiwan has typhoon today).” The results for sentence S1and sentence S3 seem confused because the readers cannot readthe summary without wondering why the typhoon is both goingaway and approaching. This is one of the common problemsassociated with many current summary generation systems;the summary sometimes contains sentences that might notbe strongly related. To solve this problem for our approach,we can set the possibility for 0.8, and then, the sentenceS3 will be filtered out by the Sentence Generator. Anotherapproach for solving this problem is to enhance the structuresof the domain ontology and fuzzy ontology. In this case, if weextract the typhoon’s name of sentence S1 “ (TyphoonLeimashi)” and sentence S3 “ (Typhoon Zhatan)”from the news, and collect them to the typhoon concept of thedomain ontology, then the news agent will generate the sentencewith the names. In this way, it may solve this confused problem.Figs. 24(a) and 25(a) show the news of rain event that retrievedfrom the ChinaTimes website and the summarization result,


(a)

(b)

Fig. 23. (a). Summarization result of Chinese weather news for typhoon event. (b) Corresponding English version of Fig. 23(a).

(a)

(b)

Fig. 24. (a) News of rain event retrieved from ChinaTimes website. (b) Corresponding English version of Fig. 24(a).

respectively. Figs. 24(b) and 25(b) show the correspondingEnglish version of Figs. 24(a) and 25(a), respectively.

Figs. 26(a) and 27(a) show the news of cold current event thatretrieved from the ChinaTimes website and the summarizationresult, respectively. Figs. 26(b) and 27(b) shows the corre-sponding English version of Figs. 26(a) and 27(a), respectively.

VI. CONCLUSIONS AND FUTURE WORK

In this paper, a fuzzy ontology and its application to newssummarization are presented. The fuzzy ontology with fuzzyconcepts is an extension of the domain ontology with crisp con-cepts that is more suitable to describe the domain knowledgefor solving the uncertainty reasoning problems. In our previouswork [8], we have presented a five-layer Fuzzy Inference Agent(FIA) [7], [10] for news event ontology extraction. In this paper,we modify the FIA and extend it to a seven-layer Fuzzy Infer-ence Mechanism (FIM) to create the fuzzy ontology. Layers 6

(a)

(b)

Fig. 25. (a) Summarization result of Chinese weather news for rain event.(b) Corresponding English version of Fig. 25(a).

and 7 of FIM are utilized to compute and integrate the fuzzymembership degrees and then generate the fuzzy concepts foreach event of the news. In addition, a news agent based onthe fuzzy ontology is also developed for news summarization.


(a)

(b)

Fig. 26. (a) News of cold current event retrieved from ChinaTimes website. (b) Corresponding English version of Fig. 26(a).

(a)

(b)

Fig. 27. (a) Summarization result of Chinese weather news for cold currentevent. (b) Corresponding English version of Fig. 27(a).

The news agent contains five modules, including a retrievalagent, a document preprocessing mechanism, a sentence pathextractor, a sentence generator, and a sentence filter, to per-form news summarization. Furthermore, we construct an experi-mental website to test the proposed approach. The experimentalresults show that the news agent based on the fuzzy ontologycan effectively operate for news summarization.

Besides, there are still some problems needed to be furtherstudied in the future. For example, the summary sometimescontains sentences that might not be strongly related. Thisproblem is also one of the common bottlenecks associated withmany current summary generation systems. The further studiesabout how to define a suitable threshold to generate the sentenceset is among our future work. In addition, although this paperpresents a new approach to solve the Chinese summarizationproblem, the measured results for precision and recall stillcould be further improved. Moreover, the structure of the fuzzyontology, including the fuzzy concepts and fuzzy relationships,will be further refined and extended to various applicationdomains. Finally, in the future, we will extend the news agentto process the English news by utilizing another part-of-speechtagger and English dictionary like WordNet.

ACKNOWLEDGMENT

The authors would like to thank the anonymous referees fortheir constructive and useful comments.

REFERENCES

[1] H. Alani, S. Kim, D. E. Millard, M. J. Weal, W. Hall, P. H. Lewis, andN. R. Shadbolt, “Automatic ontology-based knowledge extraction fromweb documents,” IEEE Intell. Syst., vol. 18, no. 1, pp. 14–21, Jan./Feb.2003.

[2] T. H. Cormen, C. E. Leiserson, and R. L. Rivest, Introduction to Algo-rithms. New York: McGraw-Hill, 1999.

[3] D. W. Embley, D. M. Campbell, R. D. Smith, and S. W. Liddle, “On-tology-based extraction and structuring of information from data-richunstructured documents,” in Proc. ACM Conf. Inf. Knowledge Manage.,1998, pp. 52–59.

[4] N. Guarino, C. Masolo, and G. Vetere, “OntoSeek: Content-based accessto the web,” IEEE Intell. Syst., vol. 14, no. 3, pp. 70–80, May/Jun. 1999.

[5] H. V. Halteren, “New feature sets for summarization by sentence extrac-tion,” IEEE Intell. Syst., vol. 18, no. 4, pp. 34–42, Jul./Aug. 2003.

[6] S. Handschuh, S. Staab, and F. Ciravegna, “S-CREAM-semi-automaticCREAtion of metadata,” in Proc. 13th Int. Conf. Knowledge Eng.Manage., 2002, pp. 358–372.

[7] Y. H. Kuo, J. P. Hsu, and C. H. Wang, “A parallel fuzzy inference modelwith distributed prediction scheme for reinforcement learning,” IEEETrans. Syst., Man, Cybern. B, vol. 28, no. 2, pp. 160–172, Apr. 1998.

[8] C. S. Lee, Y. J. Chen, and Z. W. Jian, “Ontology-based fuzzy event ex-traction agent for chinese e-news summarization,” Expert Systems WithApplications, vol. 25, no. 3, pp. 431–447, 2003.

[9] W. Lam and K. S. Ho, “FIDS: An intelligent financial web news articlesdigest system,” IEEE Trans. Syst., Man, Cybern. A, vol. 31, no. 6, pp.753–762, Nov. 2001.

[10] C. T. Lin and C. S. G. Lee, “Neural-network-based fuzzy logic controland decision system,” IEEE Trans. Comput., vol. 40, pp. 1320–1336,Dec. 1991.

[11] N. Lammari and E. Metais, “Building and maintaining ontologies: A setof algorithm,” Data Knowledge Eng., vol. 48, pp. 155–176, 2004.

[12] I. Mani, “Recent developments in text summarization,” in Proc. CIKM,2001, pp. 529–531.

[13] K. R. McKeown, R. Barzilay, D. Evans, V. Hatzivassiloglou, M. Y. Kan,B. Schiffman, and S. Teufel, “Columbia multi-document summariza-tion: Approach and evaluation,” in Proc. DUC Conf. Text Summariza-tion, 2001.


[14] R. Navigli and P. Velardi, “Ontology learning and its application to au-tomated terminology translation,” IEEE Intell. Syst., vol. 18, no. 1, pp.22–31, Jan./Feb. 2003.

[15] E. Riloff, “An empirical study of automated dictionary construction forinformation extraction in three domains,” Artificial Intell., vol. 85, no.1–2, pp. 101–134, Aug. 1996.

[16] V. W. Soo and C. Y. Lin, “Ontology-based information retrieval in amulti-agent system for digital library,” in Proc. 6th Conf. Artificial Intell.Applicat., 2001, pp. 241–246.

[17] M. Vargas-Vera, E. Motta, J. Domingue, M. Lanzoni, A. Stutt, and F.Ciravegna, “MnM: Ontology driven semi-automatic and automatic sup-port for semantic markup,” in Proc. 13th Int. Conf. Knowledge Eng.Manage., 2002, pp. 379–391.

[18] H. J. Zimmermann, Fuzzy Set Theory and its Applications. Boston,MA: Kluwer, 1991.

[19] (2003) Chinese Knowledge Information Processing Group. Aca-demic Sinica, Taipei, Taiwan, R.O.C. [Online]. Available: http://godel.iis.sinica.edu.tw/CKIP/

[20] “Chinese Electronic Dictionary, Technical Report,” Academia Sinica,Taipei, Taiwan, R.O.C., no. 93–05, 1993.

Chang-Shing Lee received the B.S. degree in infor-mation and computer engineering from the ChungYuan Christian University, Chung-Li, Taiwan,R.O.C., in 1992, the M.S. degree in computer sci-ence and information engineering from the NationalChung Cheng University, Chia-Yi, Taiwan, in 1994,and the Ph.D. degree in computer science andinformation engineering from the National ChengKung University, Tainan, Taiwan, in 1998.

He is currently an Associate Professor with the De-partment of Information Management, Chang Jung

Christian University, Tainan. His research interests include intelligent agent, on-tology engineering, knowledge management, Web services, semantic Web, andsoft computing systems. He holds several patents on ontology engineering, doc-ument classification, and image filtering.

Dr. Lee received the MOE’s Campus Software Award in 2002, the CJCU’sOutstanding Research Achievement Award in 2003, and the OutstandingTeacher Award from Chang Jung Christian University in 2004. He has guestedited a special issue for the Journal of Internet Technology. He is a Memberof TAAI.

Zhi-Wei Jian was born in Ilan, Taiwan, R.O.C., in1980. He received the B.S. degree in informationmanagement from the Chang Jung Christian Uni-versity, Tainan, Taiwan, in 2002. He is currentlyworking toward the M.S. degree in computer scienceand information engineering at National ChengKung University, Tainan.

His research interests include document sum-marization, ontology application, natural languageprocessing, and knowledge acquisition.

Lin-Kai Huang was born in Changhua, Taiwan,R.O.C., in 1981. He received the B.S. degree ininformation management from the Southern TaiwanUniversity of Technology, Tainan, Taiwan, in 2003.He is currently working toward the M.S. degree ininformation management at Chang Jung ChristianUniversity, Tainan.

His research interests include ontology applica-tions, quality management, and CMMI.

A Fuzzy Ontology and Its Application to News Summarization

Documents

Transcript of A Fuzzy Ontology and Its Application to News Summarization